Summary

• Fivetran-integrated Rocky users who need to branch logic on connector type (stock vs custom reports, schema prefix, custom tables) can now read SourceOutput.metadata.<field> directly from rocky discover output instead of re-parsing connector-type strings or re-calling the Fivetran REST API.
• Adapter-neutral metadata: IndexMap<String, serde_json::Value> field added to both rocky_core::source::DiscoveredConnector and rocky_cli::output::SourceOutput. Populated per-adapter; other adapters (Airbyte, DuckDB, Iceberg) opt in by stamping keys — no trait change required.
• The Fivetran adapter is the first populator. metadata_from_connector projects a stable, namespaced subset of the Fivetran connector config into the adapter-neutral map under fivetran.* keys.

Field shape

SourceOutput.metadata: IndexMap<String, serde_json::Value> with #[serde(default, skip_serializing_if = "IndexMap::is_empty")].

• IndexMap (not HashMap) so iteration order is insertion-stable — the discover JSON output stays byte-stable across runs, which matters for the dagster fixture corpus and the codegen-drift CI check. Mirrors the existing SourceOutput.components precedent.
• skip_serializing_if = "IndexMap::is_empty" so adapters that haven't opted in (DuckDB, Iceberg, Airbyte) produce the same wire output as before. Verified via two consecutive just regen-fixtures runs — the DuckDB playground fixture is byte-identical.
• Values are opaque serde_json::Value (not a specific struct) so Rocky relays service-specific payloads verbatim without modelling every Fivetran service. Classification is downstream.

Fivetran projection keys

Codegen cascade (just codegen)

Regenerated with the new field:

• schemas/discover.schema.json — new metadata property under SourceOutput
• integrations/dagster/src/dagster_rocky/types_generated/discover_schema.py — autogenerated Pydantic (metadata: dict[str, Any] | None = None)
• editors/vscode/src/types/generated/discover.ts — autogenerated TypeScript interface

Hand-written Python mirror

SourceInfo in integrations/dagster/src/dagster_rocky/types.py is the hand-written Pydantic model that RockyComponent / translator / sensor code consumes (parallel to the autogenerated SourceOutput). The metadata field is mirrored there as dict[str, Any] = Field(default_factory=dict) so translator/component code iterating source.metadata.items() doesn't need a None-guard. Note: the hand-written and autogenerated empty-case defaults diverge slightly ({} vs None) — the hand-written one is the intended consumer surface per the FR's downstream pattern.

RockyDagsterTranslator.get_metadata forwards adapter-namespaced keys into AssetSpec.metadata: string values pass through, richer structures are JSON-encoded so Dagster's dict[str, str] contract holds (consumers json.loads() the projected blobs as needed).

Test plan

• cargo test -p rocky-fivetran — 26 unit tests + 11 wiremock tests pass. 6 new unit tests added (3 metadata_from_connector projections, 2 end-to-end wiremock discover paths, 1 config-blob deserialization).
• cargo test -p rocky-core — 207 tests pass. New round-trip test for DiscoveredConnector.metadata serialization stability.
• cargo test -p rocky-cli — 1014 tests pass. Existing discover tests unchanged.
• cargo test --workspace — no regressions across the 20-crate workspace.
• cargo clippy --workspace --all-targets -- -D warnings — clean.
• cargo fmt --all --check — clean.
• just codegen — regenerates three files (schema, Pydantic, TS) cleanly.
• just regen-fixtures — byte-stable across two consecutive runs (same MD5 for discover.json).
• uv run pytest in integrations/dagster/ — 316 tests pass. 3 new tests added (DiscoverResult parsing with metadata, generated SourceOutput metadata exposure, translator forwarding + backward-compat regression guard).
• uv run ruff check + ruff format --check — clean.
• npm run compile in editors/vscode/ — clean TS build.

Forward pointer

Other adapters (Airbyte, DuckDB, Iceberg, Snowflake sources, custom) opt in by populating the optional field under their own namespaced keys (airbyte.*, snowflake.share_id, bigquery.labels, ...). No trait change is required — the DiscoveryAdapter::discover return type already carries the metadata map.